If you don’t care about Talos performance results, or Talos hardware, stop reading now!

In the last dev platform meeting, I mentioned we’d be doing a lot of Talos recalibration work in early January 2010, and promised updates as we figured out the plan. We’re still working out some details, but here’s what we know so far.

Talos runs performance tests on mac mini hardware. The 1.83GHz minis we use today were all spec’d 2+ years ago, when we migrated from the previous 1.66GHz minis. During 2009, we kept buying and powering up more 1.83GHz minis to keep up with increasing demand of:

• having more developers doing more frequent checkins
• adding more talos suites to run per checkin and
• running Talos on additional new OS.

However, these 1.83GHz minis have long since been discontinued by Apple, so we can’t buy any more. The “trade-in” program we ran in November got us another ~42 minis, bringing us up to 159 minis – barely enough to support the extra load needed through the FF3.6 release cycle.

At this point, our only option is to buy a whole larger set of newer spec minis, and recalibrate Talos performance results on those newer spec minis. The new spec minis are: 2.26GHz minis, with 2GB RAM and 160GB disk. We’ll recalibrate the Talos results by running the new minis concurrently with the old minis for a week or two. After we verify there are no inconsistencies in the results, we can then power off the old Talos minis, and recycle them for use elsewhere.

To avoid possible confusion about performance results, this is being done soon after FF3.6 ships and before FF3.7 works ramps up. In the meanwhile, we’re now starting the behind-the-scenes unboxing, racking, networking and configuration work.

This means a few things:

• changing Talos hardware means re-calibrating all Talos results, including results for a list of selected important historical milestones. We’re still figuring that list out.
• its important that all talos machines, across all branches, all OS, are identical, so developers can do like-with-like comparison of results across the board. This means we need to replace all the minis at one time.
• for OSX, we’ll have to change from using OSX 10.5.2 to using OSX 10.5.8. This is because 10.5.2 does not work on the new minis. This OS change *might* modify performance results so we held off doing it until this time.
• these new minis will be able to support OSX 10.6 when we add 10.6 support to Talos.
• we’re working with IT to figure out how to rack,power,network these machines here in 650castro, as the MPT colo is already full.
• we’re also using this time to move these machines from the QA network (where Talos was originally created) over to the Build network along with the rest of the build and unittest systems. We’ll also bring the logins, pswds, etc in sync to make them all easier to maintain.

Obviously, there are lots of details in this project to be careful of. If you think we’ve missed something please let us know. I’m cross-posting this to a couple of newsgroups to make sure this important upcoming change is noticed. However, please post followups in dev.planning or in bug#537065.

Thanks for reading this far, and take care
John.

(Maybe this is old news to everyone else, but I stumbled across this article recently, and found it quite interesting.)
http://money.cnn.com/magazines/business2/business2_archive/2006/07/01/8380230/index.htm

The term “mighty micro multinational” was new to me, but after reading this article from 2006, I think the description fits Mozilla fairly well. Basically, as I understand it, the writer makes the distinction between:

• traditional companies which are based in one location until forced to react to change. Examples include ExxonMobile and IBM. This change can be because of acquisition (buying a company in a different city/state/country), offshoring (moving “less critical” work to cheaper locations, while keeping “critical” work in HQ), real estate (outgrow physical HQ, and cant find adjacent office space in the neighborhood), or a whole range of other external factors.
• a “new” category of companies which are multi-location, and multi-national, from their very inception. Literally they are location agnostic. And they do this by choice. Examples include Skype, MySQL and VistaPrint; interestingly enough, Mozilla was not mentioned in the article at all.

The difference between these two types of companies is important in internal coordination, communication, attitudes to different cultures, and most importantly hiring. Being location agnostic means able to hire “the best person for the job”, not just “the best person for the job who lives nearby, or is willing to relocate for the job”. Quote: “…it doesn’t matter much where [the developer] is physically as long as he has a broadband connection.” To handle the extra complexity, the company is held together by having people who are naturally able to bridge the different cultures, keep all the communication flowing across time zones, and are skilled in detangling communication snafus whenever they arise. He calls that role “the magic expatriate”, and as far as I can tell, Mozilla has an amazing collection of magic expatriates!

Software engineers scattered around the world each with a laptop, VOIP headset and IM/IRC sounds slightly futuristic in the business-focused article, yet its normal life in Mozilla for years now. And its very cool.

(As an exercise to the reader, just try saying “Mozilla the mighty micro multinational” 3 times in a row quickly!)

Summary:

• Overall, total numbers approx the same as last month. However, it is interesting to see fewer checkins on mozilla-central, but a significant increase in mozilla-191, mozilla-192 checkins. I suspect this is caused by increased focus on security releases and the upcoming FF3.6 release. The increased load on try is probably also related to this.
• The numbers for this month are:
• 1,675 code changes to our mercurial-based repos, which triggered:
• 20,029 build jobs, or ~84 jobs per hour.
• 40,421 unittest jobs, or ~56 jobs per hour.
• 41,467 talos jobs, or ~58 talos jobs per hour.
• 18th November was our third busiest day of the year, with 103 checkins. For comparison, this high level was only exceeded on 22nd Sept (116 checkins) and 20th May (108 checkins).
• We enabled new Talos suites, and also disabled some unittest suites at different times during the month. For simplicity, I’ve ignored those changes for now, and will include in next month’s data.
• We are still not tracking down any l10n repacks, nightly builds, release builds or any “idle-timer” builds.

Here’s how it looks compared to the year so far:

Detailed breakdown is :

Here’s how the math works out:

The types of build, unittest and performance jobs triggered by each individual push are best described here.

Another round of construction started here recently in the Mountain View office.  They’re trying really hard to keep the dust and disruption to a minimum so they hung plastic sheeting over doorways, and taped plastic over the carpets in the corridors – its even inside the elevators.

Its funny how quickly you can get used to working in what feels like the set of a bad SciFi movie! However, while swiping my ID card on the way back to my desk, the following made me stop, double-take and then start carefully looking around me.

“Demo: Not now”

To be clear, in this context:

demo != demonstration
demo == demolition

Turns out, the entire corridor I was standing in was going to disappear… just not now.

Some of you may have noticed this new item on this menu on GraphServer.

There’s been a lot of work with linux64 over the last few weeks behind the scenes.

1) There are now nightly and per-checkin builds available for mozilla-central, mozilla-192, mozilla-191, tracemonkey, electrolysis. Because we only have 10 linux64 build slaves, we dont have builders on Places, TryServer or the cvs-based mozilla-190/Firefox3.0.

2) We’ve got a pool of linux64 talos slaves running all the usual Talos suites, per build, on those same branches. You can now see those results on graphs.mozilla.org, listed just like any other OS. Just like it should be. 🙂

3) Caveats:

• For the sake of speed, we’ve cloned the *one* preexisting linux64 machine (which dbaron? setup up), without generating a clean, new, refimage with fully identified toolchain. If you see any toolchain problems, please let us know, but as its identical to whats been in place before, hopefully it will continues to be good enough for now.
• Unittests are not yet being run on linux64. This is being worked on as part of a bigger problem; unittests used to require doing a build first. This in turn meant we only could run unittests on platforms that we supported using for builds, so we dont have unittests on 10.4, xp, vista, etc. More on this as it develops, but its not complete yet.
• We’re still working out some TBPL display updates to get linux64 showing up on TBPL. For now, you must use Tinderbox waterfall to see the linux64 builds. The curious can follow bug#532560)

Spinning up this new OS took work from most people in the group, and is the first new desktop platform we’ve supported in years. Very very cool work and a great way to end the week. Enjoy!